Uploaded September 2026 | Updated November 2026, 4 days ago
Akshay Sharma, Machine Learning Engineer at Lyft, breaks down how Lyft rebuilt its customer support agent stack with LangGraph and LangSmith. The team replaced brittle, single-call agents with a meta-agent architecture where PMs and ops can ship new agents as a prompt and a config change, without writing new code. The result: agent build time dropped from six months to one to two weeks, and resolution rate is up 16%.
Chapters:
0:00 The scale of Lyft's customer support problem
0:23 The metric that matters: resolution rate
0:41 Before LangSmith: basic, deterministic agents
0:59 The observability blind spot
1:32 Engineers stuck building agents instead of solving hard problems 1:45 The fix: a self-serve platform for PMs and ops
2:15 Inside the architecture: a meta-agent with LangGraph nodes
2:44 LangSmith tracing and evals: from black box to visibility
2:56 The result: 6 months to 2 weeks, and a 16% lift in resolution rate
Resources:
→ LangGraph: langchain.com/langgraph
→ LangSmith: langchain.com/langsmith
→ LangChain Academy: academy.langchain.com
Akshay Sharma, Machine Learning Engineer at Lyft, breaks down how Lyft rebuilt its customer support agent stack with LangGraph and LangSmith. The team replaced brittle, single-call agents with a meta-agent architecture where PMs and ops can ship new agents as a prompt and a config change, without writing new code. The result: agent build time dropped from six months to one to two weeks, and resolution rate is up 16%.
Chapters:
0:00 The scale of Lyft's customer support problem
0:23 The metric that matters: resolution rate
0:41 Before LangSmith: basic, deterministic agents
0:59 The observability blind spot
1:32 Engineers stuck building agents instead of solving hard problems 1:45 The fix: a self-serve platform for PMs and ops
2:15 Inside the architecture: a meta-agent with LangGraph nodes
2:44 LangSmith tracing and evals: from black box to visibility
2:56 The result: 6 months to 2 weeks, and a 16% lift in resolution rate
Resources:
→ LangGraph: langchain.com/langgraph
→ LangSmith: langchain.com/langsmith
→ LangChain Academy: academy.langchain.com










